Feature Selection Active Learning Algorithm Based on Sparse Subspace

Feature Selection Active Learning Algorithm Based on Sparse Subspace

Active learning has been widely used in many fields, and has proved to be very effective to solve a vast array of machine learning tasks, but it still has some shortcomings. The problem is that it only pays attention to label information of samples and ignores the structural information, so it cannot reach state-of-art performance in some tasks and is sensitive to initial data. In this paper, a new algorithm is proposed by combining Sparse Subspace algorithm with active learning together. Firstly, Sparse Subspace algorithm is used to find the dataset’s structure. Then, an active learning algorithm based on feature selection is adopted. In this way, not only the label information but also the distribution information of samples is taken into account. As a result, the efficiency and robustness of active learning are improved, and the labeling cost of samples and the probability of falling into local optimum are reduced.